Flood control and water management have always been critical challenges for communities around the world. As climate change brings more frequent and severe weather patterns, the need for innovative solutions to predict, mitigate, and respond to flooding has never been greater. One promising approach that is transforming the field is the integration of machine learning (ML) algorithms into flood forecasting systems.
Now, this might seem counterintuitive…
Advancing Stage Forecasting with LSTMs
At the heart of an effective flood forecasting system is the ability to accurately predict water levels, or river stage, at key locations along a river network. Traditionally, this has been done using physics-based hydrological models that simulate the complex interactions between rainfall, soil moisture, and river flows. However, these models often struggle with capturing the nuances of local watersheds and can be computationally intensive to set up and maintain.
In contrast, deep learning-based approaches, such as Long Short-Term Memory (LSTM) networks, have shown remarkable promise in improving the accuracy and scalability of stage forecasting. LSTMs are a type of recurrent neural network that can effectively learn long-term dependencies in time series data, making them well-suited for modeling the complex dynamics of river systems.
Google’s operational flood forecasting system, for example, uses an LSTM-based stage forecasting model that takes in past river stage measurements, precipitation data, and other relevant inputs to predict future water levels. These models have demonstrated significantly higher skill compared to traditional linear regression approaches, with median Persistent-NSE scores of 0.69 across hundreds of target gauges in India and Bangladesh.
The LSTM stage forecasting models are also designed to be regionally calibrated, meaning a single model can be trained on data from multiple basins and then fine-tuned for each individual gauge. This allows the models to effectively leverage shared patterns in hydrological behavior while still capturing the unique characteristics of a given river system.
Enhancing Inundation Modeling with Machine Learning
Once accurate stage forecasts are obtained, the next step is to translate these water level predictions into flood inundation maps that delineate the expected extent and depth of flooding. Traditionally, this has been done using physics-based hydraulic models that numerically solve the Saint-Venant equations governing open-channel flow.
However, these hydraulic models can be computationally intensive, requiring detailed bathymetric data and extensive calibration. To address these challenges, researchers have been exploring the use of ML-based inundation modeling techniques.
One innovative approach, pioneered by Google’s flood forecasting team, is the “morphological” inundation model. This model combines the strengths of physics-based and data-driven approaches by learning a direct mapping between river stage and flood extent from historical data, while still enforcing physically reasonable constraints on the resulting inundation maps.
The morphological model first learns a stage-to-profile relationship that predicts the one-dimensional river profile from a given stage measurement. It then uses this learned profile to modify the underlying digital elevation model (DEM) in a way that replicates the expected flooding pattern, without the need for computationally intensive hydraulic simulations.
Compared to traditional hydraulic models, the morphological approach has been shown to improve inundation mapping accuracy by up to 3% while requiring significantly less manual effort and computational resources. This makes it possible to scale the inundation modeling component to cover large areas efficiently.
Expanding Flood Alerts through Scalable ML
While the stage forecasting and inundation modeling components are crucial, the ultimate goal of a flood forecasting system is to provide timely and effective alerts to the affected population and emergency responders. This requires not only accurate predictions, but also the ability to reach a wide audience with relevant, actionable information.
To address this challenge, the Google flood forecasting system employs an additional “alert targeting” model that uses publicly available data, such as stream gauge measurements and low-resolution satellite imagery, to directly infer flood inundation patterns. This end-to-end ML approach allows the system to expand coverage to areas not served by the more detailed morphological inundation models, dramatically increasing the number of people who can receive flood alerts.
During the 2021 monsoon season in India and Bangladesh, the Google flood forecasting system was able to provide over 100 million flood alerts to the affected population, reaching an estimated 22 million people. These alerts included not only the forecasted water level changes, but also detailed inundation maps and depth information where available, empowering communities to take appropriate preparedness and response actions.
The Future of Flood Forecasting
As the field of flood control and water management continues to evolve, the integration of machine learning algorithms is poised to play an increasingly pivotal role. By enhancing stage forecasting, inundation modeling, and alert targeting, these innovative techniques are helping to improve the accuracy, lead time, and reach of flood forecasting systems.
Looking ahead, researchers are exploring even more advanced ML architectures, such as “HydroNets”, which leverage the underlying structure of river networks to further improve forecasting performance and generalization across different basins. These next-generation models hold the promise of delivering flood forecasts with unprecedented accuracy and lead time, ultimately saving lives and reducing the devastating impacts of flooding around the world.
To learn more about the latest developments in flood control and water management, be sure to visit Flood Control 2015. This comprehensive resource covers the design, implementation, and maintenance of a wide range of flood control systems, including levees, flood barriers, storm drainage, and emergency response strategies. Stay informed and join the conversation on innovative flood forecasting and mitigation techniques.
Tip: Regularly inspect and maintain flood barriers and drainage systems
